fix(models): keep adaptive thinking for Claude 5 and default Haiku to 5.5

[NanmiCoder]

Summary

Requests to claude-opus-5-5 through Modelflare failed with 400 claude-opus-5-5 requires adaptive thinking. With CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 set, the CLI sent thinking: {type: "enabled", budget_tokens}. Two things combined to cause this: the provider capability string for Claude models omits required_thinking, and the env var fell back to a manual budget. Some relays rewrite the field (AruHub), but strict ones forward it and the API rejects it.

Per the official thinking table, budget_tokens returns 400 on every Claude 5.x model and on Opus 4.7/4.8. The official CLI removed this env var in 2.1.260.

• These Claude models now always use adaptive thinking, whatever the env var says. When a provider's capability list declares adaptive_thinking but leaves out required_thinking, adaptive-only models still stay adaptive. A provider can still opt out by leaving out adaptive_thinking.
• DeepSeek, MiniMax, GLM, Kimi, Doubao, Hunyuan and Claude 4.6 and earlier keep sending the manual budget, as before.
• Haiku 5.5 is now supported:
  • its own canonical name (it previously collapsed to claude-haiku)
  • 1M context and 128K output
  • all effort levels, default medium
  • adaptive thinking; when thinking is disabled, xhigh/max are capped to high
  • pricing tiered by prompt length (over 100K tokens costs more) in both cost tables, with the local index bumped to v11 so stored dollars are recomputed
• The first-party Haiku default is now Haiku 5.5, matching the official model list and prompt lineup. Third-party providers keep the lag-safe Haiku 4.5 default.

Feature Quality Contract

• Changed surface: provider-runtime, server (models API, official runtime, local index version), desktop model catalog
• Tests added or updated:
  • request-shape matrix of 16 models through a provider-managed environment with the env var set. Before the fix, all 8 Claude 5.x/4.7/4.8 cases failed and the 8 non-Claude/older cases passed. After the fix, all 16 pass.
  • Modelflare-shaped regression where a fake upstream rejects anything except adaptive
  • Haiku 5.5 effort cap when thinking is disabled. Mutation-checked: removing the condition makes the xhigh and max cases fail.
  • haiku55.test.ts: identity, limits, effort, thinking, prompt-length pricing in both tables, alias defaults
  • pricing parity entry; model list/default expectations updated
• Coverage evidence: not run (check:coverage)
• E2E / live-model evidence: not run; no live provider was used
• Known risk / rollback: changes the default Haiku for first-party users. Revert this single commit to roll back.

Verification

Run on the rebased HEAD (target origin/main b6388b4):

• bun run check:server: 616 files, 7405 passed, 0 failed
• bun run check:provider-contract: 36 suites passed
• focused model/thinking/pricing suites: 206 passed
• desktop vitest (model selector, catalog, provider store, settings, agent teams): 19 files, 442 passed
• git diff --check: clean
• not run: check:agent-e2e, check:persistence-upgrade, check:coverage, bun run verify

[github-actions]

PR quality triage

Changed areas: area:cli-core, area:desktop, area:server

CLI core policy: CLI-core change authored by a CODEOWNER; no override label needed.

Missing-test policy: No missing-test policy block detected.

Coverage baseline policy: No coverage-baseline policy block detected.

CLI core files:

• src/utils/betas.ts
• src/utils/commitAttribution.ts
• src/utils/context.ts
• src/utils/effort.ts
• src/utils/model/agent.test.ts
• src/utils/model/configs.ts
• src/utils/model/fable.test.ts
• src/utils/model/haiku55.test.ts
• src/utils/model/model.ts
• src/utils/model/modelContextWindows.ts
• src/utils/model/modelOptions.ts
• src/utils/modelCost.ts
• src/utils/modelCostParity.test.ts
• src/utils/thinking.ts
• src/utils/usageAccounting.ts

Coverage policy files:

• none

Required check plan:

• See PR Quality / scope-plan; it is the single source of truth for selected jobs.

Test coverage signals:

• Agent/model runtime path changed: use mock/request-shape tests in PR and maintainer live-model smoke before release.

Risk notes:

• No special risk notes from changed paths.

Hard merge gates come from the deterministic GitHub Actions contract lanes above.